I am the shape the water takes.
— ClawdBot, Moltbook

TL;DR:

● Scaffolding — the software environment and contextual documents provided to an AI model at deployment — can yield significant performance improvements. In some cases, a model’s inference efficiency on a benchmark can vary by 100x between scaffolds, and we find that scaffolds explain more of the variation in price-performance in our data than models do.

● Unlike many ML innovations, the same scaffold can have different effects with different models and on different tasks: some models see large benefits from a given scaffold — while others see little advantage or may even be hindered.

● These scaffold-model interactions have important implications for performance, AI evaluation and the AI agent economy. We speculate that they may be a driver of increased concentration in the AI industry.

When we think about what drives AI progress, we usually point to pretraining, reinforcement learning, and inference-time efficiency. Scaffolding gets far less attention, even as it plays a growing role in how systems actually perform. If we want a full picture of what drives progress in language models, we need to understand scaffolding too. Scaffold effects can be large enough that in some cases it makes nearly as much sense to ask which scaffold you’re running as which model.

Scaffolds — also called model wrappers or harnesses — are the software programs that turn an AI model (or AI models, plural) into an agent. This post tries to get a handle on the effects of scaffolding using data from the Holistic Agent Leaderboard (HAL). HAL is an index of model performance on agentic benchmarks, reporting both accuracy and cost. In the rest of this post, we’ll:

1. Define “scaffolding” in a bit more detail

2. Describe the data we use

2. Describe scaffolding’s impact on model performance and evaluation

3. Discuss the significance of our findings for the agent economy and model security.

What is Scaffolding?

Scaffolds are software and tooling that AI models interact with (or interact within), designed to improve AI operations or unlock new capabilities. Loosely, what we have in mind is software that manages stuff like:

• Tool affordances: Bash terminal access, file editing tools, search tools, ability for one model to call other models, API calls, etc.

• Context window management: when is RAG used, how is compaction done?

• Memory management: When and where content is stored and removed

• Prompting: System prompts, planning prompts

• Pre-written skills documents (like Claude Skills), which prompt the model with guidance on how to perform specific types of tasks

• The calling of other models: How often other models are called, for what purpose, how many, etc

• Best of k sampling, majority voting, and other aggregation procedures

• Reflection, summarization, and self-critique steps

• Token time cost limits, stopping rules

• Agent loop: (for example) think → act → observe → repeat

There is some previous work analyzing the effect of scaffolding. For example, Davidson et al. (2023) examined the impact of some of “post-training enhancements,” including scaffolding, finding that some techniques yield performance improvements equivalent to a 10-100x increase in pre-training compute. For instance, they find that the LATS agent scaffold (Zhou et al, 2023) yields an approximately 10x compute-equivalent gain.

However, agentic scaffolding has become significantly more complex since 2023. In this post, we try to assess the current contribution of scaffolds to AI agent performance. And we hope future work will update ours as models and scaffolds advance.

Further work would also hopefully be able to disentangle the relative importance of the scaffold components we mentioned above. Unfortunately, we do not have sufficiently fine-grained data to distinguish the contribution of these components in present-day systems.

Our Data and Methods

We consider this post a first pass at quantifying the impact of scaffolds. Accordingly, the data analysis to generate the plots was conducted and reviewed by Claude Opus according to our guidance (the code is available on Github here).

To quantify the performance effects of scaffolding, we use data from the Holistic Agent Leaderboard (HAL). HAL is an index of model performance with multiple scaffolds on agentic benchmarks, reporting both accuracy and cost. 18 models are included — frontier models spanning a range of release dates produced by Anthropic, OpenAI, Google DeepMind, and DeepSeek (full model table in Appendix).

HAL bench includes many agentic benchmarks including CORE-Bench Hard, GAIA, Online Mind2Web, SWE-bench Verified Mini, SciCode, ScienceAgentBench, TAU-bench Airline, and USACO. These cover areas in software engineering, browser use, agentic search, scientific reasoning, and multimodal vision reasoning. (For a description of each of these benchmarks and what tools / skills they require, see the Appendix.) What makes it useful here is that it runs each model under multiple scaffolds per benchmark — typically a benchmark-specific scaffold alongside a general-purpose one. (Although it’s important to note that, despite being benchmark-specific, many of these “specialist” scaffolds are quite general in their design and affordances.) On CORE-Bench, for instance, models run under both the CORE-agent scaffold and the HAL generalist scaffold. The generalist scaffold is built on the smolagents framework: the model takes every action by writing Python code, so it “searches the internet,” for example, by writing and running code that performs the search.

Most of our data consists of specialist scaffolds designed for particular benchmarks; we also have data on a handful of generalist scaffolds, including Claude Code. Comparing among these scaffolds gives us a sense for the degree to which scaffolds can matter, and it provides tentative evidence that specialist scaffolds can outperform this particular generalist scaffold. We’d be excited to see future work extend this with more data (even new experiments) that use a wider range of scaffolds.

What Can This Data Tell Us?

In a lot of this analysis, we are comparing the effect of a small sample of scaffolds to a standard, but very specific scaffold — HAL. This means that many of our results are entirely relative to the choice of HAL as the baseline. For instance, if we had instead used a baseline that did not allow for tool calls, many of the scaffolds we assess would look comparatively amazing.

Another thing to note: there isn’t enough longitudinal data here to make empirical claims about how scaffolds might have improved over time or how much room for further improvement is possible.

Scaffolding Has a Large Impact on AI Performance

Scaffolding engineering is still a nascent field, so it’s not surprising that scaffolding’s effects are large and vary widely. Some specialist scaffolds provide large performance and efficiency advantages over generalist scaffolds, while others do not. To examine this, we plot each model’s final (logit transformed) benchmark score against the API cost needed to reach that benchmark score, across scaffolds in the figures below. We use the logit transform because benchmark accuracy is bounded between 0 and 1: a one-point improvement near saturation, say from 94% to 95%, often represents a larger gain in underlying capability than a one-point improvement near the middle of the scale, and the logit transformation helps adjust for this.

Below are per-scaffold price-performance frontiers for two benchmarks (CORE-Bench Hard and SciCode). We include graphs for the rest of the benchmarks in the appendix. Scaffold choice moves performance more than one might expect: in some cases, switching from a generalist to a specialist scaffold can buy ~100× cheaper performance at the same accuracy. How big is this number? For context, algorithmic progress in AI inference typically cuts prices ~10× per year at fixed performance (see our paper) — so switching scaffold can, in the strongest cases, be worth roughly two years of model progress.

Also, we see that the Claude Code scaffold with Claude models (the orange group in CORE-bench) leads to exceptionally high performance. We will speculate as to why this might be later in the piece.

The scaffold-switch vectors below trace the same model moving between two scaffolds — each arrow points from one scaffold’s (cost, accuracy) to another’s, colored by whether the switch made the model more/less accurate and more/less expensive. We display these plots for GAIA and SciAgentBench here, but include the graphs for all benchmarks in the Appendix.

We see here that scaffolding can have significant interaction effects with models, most visibly on GAIA: switching from the HAL generalist scaffold to the HF Open Deep Research scaffold improves some models while hurting others. This contrasts with most algorithmic progress in AI, which acts as a rising tide that sometimes lifts all boats — think of GeLU or flash attention, for example, where essentially every model benefits similarly from the new innovation. It’s also a contrast to ScienceAgentBench, where the ScienceAgentBench self-debug scaffold provides a more uniform benefit to all models.

It’s not entirely clear why some models are more advantaged by a given scaffold than others. One explanation could be that some scaffolds are more advantageous at low inference budgets, while other scaffolds could be better to use at higher inference scales. Interestingly, we see some evidence of this in the Pareto frontier graphs in the Appendix.

How often do scaffolds have heterogeneous, model-specific effects rather than producing a uniform, “rising tide” improvement to all models? To get a better sense, let’s look at a transformed version of the switch vector graphs we showed above. In the versions below, we center each vector on the origin, which now each represent a model’s performance on the HAL generalist scaffold. The arrows point to the difference in price and performance achieved after using alternative scaffolding — in most of the plots, specialist scaffolding adapted to the particular benchmark.

Averaging the vectors for each benchmark, we can arrive at a general picture of how scaffolding affects model performance. Overall, the majority of the specialist scaffolds measured in HAL improve the accuracy. However, some are only able to do this by using significantly more tokens. And, looking at the averages across benchmarks, there is significant heterogeneity. (Also, one caveat: it’s possible that the specialist scaffolds are “overfit” to their corresponding benchmarks, and might not generalize very well.)

Scaffolding Effect on AI Leaderboard Ranking

People sometimes create leaderboards of models based on their agentic capabilities. However, in addition to other known issues with leaderboard robustness, leaderboards can be very sensitive to scaffolding choices. If scaffolds have non-uniform effects, a model ranked 20th under one scaffold could place 3rd under another.

We include a formal analysis of scaffold effects on ranking in the appendix. On some benchmarks — SciCode, for instance — rankings are largely preserved across scaffolds (high rank correlation). But on most, rank-preservation measures show substantial reordering: the leaderboard you get depends heavily on the scaffold you chose to run.

Given this: how should leaderboard designers proceed? If the goal is to assess how models will perform under typical deployment conditions, the leaderboard should fix the scaffold that will actually be used. If the goal is to assess an AI system’s maximum capabilities under ideally conducive conditions, the leaderboard should run models under several scaffolds and take the best.

Potential Implications

Scaffolds could have a number of important implications for regulators, model developers, and society. We want to discuss a few particularly important potential implications of our findings here, as well as gesture beyond our immediate findings at future research directions.

Scaffolds May Be a New Vector of AI Capabilities Improvement

Our study here shows that scaffolds can have large impacts, and some existing scaffolds vary widely in their utility. We may soon see more effort to identify which scaffolds (and features of scaffolds) are most performant, and competition to build ever-better versions. If there is a lot of room for new innovation in scaffold designs — especially AI-automated innovation — it will become important to track progress in scaffolds alongside progress in raw model capabilities.

On the other hand, it may be that there is limited room to build scaffolds much better than we have now. If that’s the case, we should expect short term gains as the field converges to the best scaffolds of the designs available or the best models, but little progress after that.

Speculatively, one reason to doubt continued progress is this: more powerful future models may no longer need advanced agentic scaffolding. (Perhaps these future AIs will just design the tools they need right as they need them?) But our personal view is that this is less likely; we expect scaffolds’ importance will continue — or grow.

Scaffolds May Affect AI Developers’ Monopoly Power

It’s widely contested whether AI development is a natural monopoly — will big players with lots of compute, like OpenAI and Anthropic, increasingly dominate the market for AI services, or will small developers increasingly undermine their business? While we can’t say for certain, scaffolds may be an important factor in shaping the eventual equilibrium. Our results show that using good scaffolds can yield big improvements in price and performance. Accordingly, it is conceivable that scaffolds could erode the advantage of expensive frontier models, by allowing cheap models to offer improved performance — especially if scaffolds are cheap to produce.

However, we believe that the opposite outcome is more likely: scaffolds will lead to greater concentration. Why? First, while our data isn’t sufficient to show this, it seems likely that scaffolds are to some degree complementary to compute investment. If scaffolds help more expensive models more than cheaper ones, they could increase returns to scale, which would drive concentration.

Second, we expect that scaffolds will become increasingly expensive, as firms expend inference compute to optimize their designs, and scaffolds increasingly deploy complicated ensembles of models to perform more complex, long-horizon tasks. (We’d love to see more research on this question: to what extent can scaffolds be improved with more inference compute?)

Third, and perhaps more interestingly, we speculate that there may be benefits to co-optimizing models and scaffolds. For instance, Anthropic can do reinforcement learning that trains Claude within their Claude Code scaffold. And they can simultaneously adjust their scaffolding to take advantage of Claude models’ unique capabilities — which only Anthropic has white-box access to. Both the Claude models and the Claude Code scaffold might then improve together, generating a synergistic performance advantage that outside actors can’t easily replicate. It requires a prohibitive cost to train a foundational model from scratch on a proprietary scaffold. While our analysis here isn’t sufficient to demonstrate that Anthropic is already getting benefits from co-optimization, we can’t rule it out — and we do observe that Claude Code seems to perform exceptionally well.

If this sort of co-optimization becomes important, Claude models might then become hyper-optimized for the Claude scaffold, with limited interoperability with other scaffolds and downstream applications. Downstream scaffold developers without control over models might even see their scaffolds’ performance decrease rather than increase with new Claude models (see how the SWE-agent harness decreases performance for some models). Just as Apple has been able to preferentially benefit its own applications over other providers’ applications on its App Store, Anthropic could design Claude’s training to prioritize the downstream frameworks it wants to succeed.

Task-Specific Scaffolds May Become (More of) a Thing

We find numerous instances of specialist scaffolds, tailored to particular benchmark tasks, which outperform a generalist scaffold on that task. We don’t have many generalist scaffolds in our data, so it’s far too early to say that task-specific scaffolds are typically better than generalist ones, or anything like that. But it will be interesting to watch the field develop – if specialist scaffolds can outperform generalist ones, we may see companies spring up to provide task- or industry-specific scaffolds. (Arguably, this is a good description for some companies we already see, like Harvey and Cursor.)

Scaffold Security is Important

For those concerned about leakage or cybertheft of advanced AI capabilities, our findings here suggest that it may be important to keep scaffolding secure — perhaps just as important as models themselves. However, scaffolding security is significantly less discussed than model weight security. Notably, the scaffolding behind Anthropic’s breakthrough product Claude Code was accidentally leaked in March.

How much scaffold security matters depends significantly on how closely models and scaffolds are co-optimized. If scaffolding is highly tailored to a specific model, then exfiltration of either the scaffolding or the model weights alone is less severe because both scaffold and weight exfiltration are necessary to reach full performance. Companies can take active measures to prevent scaffold exfiltration being independently useful to attackers. For instance, by making the scaffold dependent on external APIs or tools that are only available at AI labs or legitimate organizations. On the other hand, if scaffolds are both powerful and useful regardless of what models they are used with, then scaffolding exfiltration might be a great risk. In this regime, a leaked scaffolding would be able to significantly improve capabilities for many unintended models at once with minimal adaptation.

Conclusion and the Future of AI Capabilities

The systems models are embedded in seem to be an important determinant of model capabilities. It’s possible that this could become increasingly so in the future. Complicated scaffolds, multi-agent networks with complex information sharing, and new structures we haven’t yet imagined may continue to blur the line between what is and is not an AI. Understanding the capabilities — as well as the alignment profile — of these new systems will then require looking beyond individual models in isolation.

Appendix

This appendix includes the following:

• A table of the benchmarks and agents in our dataset

• A table of the models in our dataset

• An analysis of the explanatory power of scaffolds vs models

• The full suite of Pareto analysis plots, vector analysis plots, and ranking correlation plots, and an additional plot showing how specialists scaffolds compare to HAL

• A link to our code and data

Benchmarks

Details of each benchmark and scaffold in our dataset. This information was compiled by Claude to our specifications.

Models in Dataset

These are all the model configurations we use, along with the number of benchmarks and scaffolds each appears with. Some models were run both at a high inference setting and with no specific setting; where both occur on the same benchmark and scaffold, we keep the higher-accuracy run. A couple of models (o3 and GPT-5) appear only at a medium setting. The exception is o4-mini, which was run at a high and a low setting (rather than high/none): we keep both.

Explanatory Power of Scaffolds vs Model Algorithmic Efficiency

Methodology: our goal here is to measure how much of the variance in (logit transformed) performance is explained by scaffolds and models, controlling for log price and benchmark. (Given our sample size, we add only an additive fixed effect for each benchmark.) The issue is that benchmarks and scaffolds are highly correlated since most of the scaffolds are used exclusively on a single benchmark; we don’t want to attribute performance changes to scaffolds that are really just differences between benchmarks, but a naive regression would have no way of distinguishing. Our fix is to build up the regression incrementally. We first regress logit performance on log price and benchmarks, and observe the adjusted R^2 (which is 0.519). Then, we refit the regression twice: once with model dummies added in, and once with the scaffold dummies added in instead. In each case, we observe how the adjusted R^2 changes compared to the baseline regression with just benchmark dummies and log price. The plot above shows the explanatory power of models vs scaffolds. The takeaway is that scaffolds explain more of the variance in our data than models do, underscoring the importance of scaffolding.

(Running a similar regression with both scaffold and model incorporated sequentially gets similar results, regardless of the order in which the parameters are incorporated into the regression. This is because there is relatively little shared variance between scaffolds and models.)

Pareto Analysis Graphs For All Benchmarks

Per-scaffold price–performance frontiers across all benchmarks. Each panel plots final accuracy (logit-scaled) against the API cost needed to reach it, with one frontier per scaffold.

Full Vector Analysis Graphs

Scaffold-switch vectors for the remaining benchmark/scaffold pairs. Each arrow traces a single model moving between two scaffolds, from one scaffold’s (cost, accuracy) to the other’s, colored by whether the switch made the model more/less accurate and more/less expensive, as in the figures in the main text. As in the main text, we see two types of scaffolding changes: rising-tide changes and turbulent changes. We see a roughly equivalent number of turbulent changes as rising-tide change

Ranking Correlation Graphs

How well does a model’s ranking on one scaffold predict its ranking on another? The chart below (left) shows Spearman ρ and Kendall τ rank correlations between scaffold pairs for each benchmark. Most pairs preserve ranking moderately well, but CORE-bench Hard (CORE-Agent → Claude Code) is notably negative — the best models on one scaffold are not the best on the other. Zooming in on GAIA (right) with τ=0.17, only about a 58% chance that any two models keep their relative order across scaffolds (this is what Kendall τ rank measures). Points on the green diagonal kept their ranking; the reds (e.g. DeepSeek V3, o3 Medium) jump far off it.

Both measures run from −1 (rankings exactly reversed) through 0 (no association) to +1 (identical rankings). Spearman’s ρ is the Pearson correlation applied to the models’ rank positions rather than their raw scores: take each model’s rank under scaffold A and its rank under scaffold B, and measure how linearly those two sets of rank numbers track each other. Because it works on the rank values, ρ is sensitive to how far a model moves — a model that falls from 2nd to 18th drags ρ down more than one that slips from 2nd to 4th. Kendall’s τ, by contrast, only counts whether each pair of models keeps the same order, ignoring the size of the move.

Specialist Scaffolding Increases Across Benchmarks

Code and Data

All data was scraped from the publicly available HAL leaderboard. The code for generating graphs and doing analysis is here: Github.

Thank you to Alex Fogelson for reading a draft of this piece and offering helpful comments.

Subscribe now

Today’s AI developers have access to a number of resources that the would-be AI-builders of yesteryear lacked:

1. Better algorithmic / architectural ideas,

2. More training data,

3. More efficient computing hardware,

4. And more cash to spend on that computing hardware and data.

Growth in each of these inputs has been necessary to reach modern AI capabilities. But growth has been historically uneven, and as a result, AI progress has been bound more by some inputs than others.

We’ll argue that hardware efficiency in particular has been the binding constraint on model capabilities over the last 70 years. The steady pace of hardware efficiency progress – described by Moore’s law – only recently brought us to the point where rapid investment in hardware quantity could close the last mile, which we’ve seen over the last decade. That investment would be for naught without the prior half-century of hardware efficiency progress.

Going forward, it’s reasonable to expect that the steady march of hardware efficiency progress will continue. But the rapid pace of investment we’ve seen recently cannot continue indefinitely unless it leads to dramatic and sustained economic growth. If transformative AI – transformative enough to bring about that kind of growth – is achievable with only a few more orders of magnitude of investment, then… buckle up! But if it isn’t, AI investment will eventually need to slow to an economically sustainable growth rate, commensurate with the returns on investment. In that scenario, AI progress may once again depend heavily on the slower pace of hardware efficiency progress.

In this post, we’ll examine the history of AI progress and assess the role of each of the four inputs. We’ll then explain why the next ten years of AI progress will be an extremely important signal for AI’s future trajectory.

This post is based on a talk delivered by Dr. Neil Thompson at the AI for the Economy Forum, which FutureTech co-hosted with Google last month. You can view that talk here.

We had the ideas, but not the compute

A typical history of an area of science emphasizes the importance of ideas. Perhaps surprisingly, though, many of the foundational ideas in AI were conceived of in the 1950s and 60s, but lacked sufficient resources to be implemented. In other words, algorithms were probably not the binding constraint on AI progress.

In a seminal 1948 paper, Claude Shannon, the founder of information theory (and the namesake of Anthropic’s Claude model), showed that it was possible for statistical measures to predict the next “token” of English language text given the preceding words. In the very same paper, he established the concept of entropy as a metric for evaluating the predictive performance of such models – the foundation of the measures we use today. Just a few years later in 1957, Frank Rosenblatt would invent the perceptron – the fundamental unit of modern deep learning. Modern systems combine millions of these digital neurons into networks, requiring massive amounts of compute to train, but even then researchers were already starting to scale perceptrons into larger neural networks, given what their available hardware would allow.

Of course, some key algorithms were admittedly still missing, but important precursors were often already around. Rosenblatt himself had already discussed the possibility of “deep” networks as early as 1962. And while backpropagation – an essential algorithm for training deep networks – would not enter general use in AI until the mid-1980s, early versions of backpropagation were actually already discovered and discussed numerous times before then.

The existence of so many essential ideas in the middle of the twentieth century lends support to the idea that algorithms were ahead of what was implementable at the time, given the hardware and training data available.

That implementation took time. It would take another 30 years before neural networks and text prediction could be connected and put into practice. In 1988, IBM created Candide, a statistical machine translation system which was trained on the Canadian Hansard, a digitized text corpus just large enough to train an n-gram model (a phrase coined by Shannon). Just two years later, Jeffrey Elman proposed using recurrent neural networks to model the next token of text..

Algorithms and data are downstream of compute

We don’t want to downplay that ideas like Elman’s required ingenuity, or suggest that ideas since the 60s have been trivial or unimportant for AI progress. We only want to suggest two things: first, that a lot of the essential ideas were floating around much earlier than they were eventually implemented and validated in AI systems. And then second: if more computing power and training data had been available for researchers in the past, the remaining algorithmic insights needed to reach today’s frontier might have been discovered even sooner. Why? Impressive results would have made it clear earlier which research directions were most promising, incentivized algorithms that take advantage of raw computational power, drawn new researchers into the field, led to earlier real-world usage of neural networks, and attracted more investment.

This idea, that lack of computing capacity constrained innovation, is not just our modern, retrospective diagnosis. Writing in 1990 on the previous thirty years of progress in neural networks, Bernard Widrow and Michael Lehr emphasize that it was indeed the available hardware that constrained the models researchers could build: “Today, most artificial neural network research and application is accomplished by simulating networks on serial computers. Speed limitations [a result of compute limitations] keep such networks relatively small [...]”

Just as algorithmic ideas might have matured sooner if more hardware was available for experiments, the massive digital text datasets we now have are also downstream of progress in hardware. In particular, internet scale data requires the internet, a massive population of internet users contributing text, and sufficiently cheap, powerful hardware (with inexpensive high-capacity memory) for ubiquitous personal computing in developed countries.12

We’ve argued that compute was the binding constraint. But was it lack of investment in compute – could early AI researchers have simply purchased more hardware? Or was it the efficiency of the hardware on offer?

The latter. To see why, let’s pose a counterfactual thought experiment: how much would someone like Claude Shannon – one of the greatest computer scientists of the mid-twentieth century – have needed to spend to train an AI model of today’s capabilities with the hardware of his time?

The cost of Claude in 1946

Today’s frontier models, like Claude Opus, require massive amounts of compute – around 10²⁶ FLOPs. Throwing aside the need for new algorithms and data, let’s look at what it would take to build such a thing with the hardware of Shannon’s era – the mid-forties.

In 1946, the Electronic Numerical Integrator and Computer (ENIAC) was built for the United States Army – the first programmable, electronic, general-purpose digital computer. ENIAC was a landmark, Turing-complete, fully reprogrammable machine. It cost the equivalent of $7 million in 2024 and operated at about 500 FLOPs per second. We’ll use ENIAC to ground our first estimates – full calculations are detailed at the end of this post.

Since we know ENIAC’s power consumption, historical pricing, and cost of construction, we can back out an estimate for dollars per FLOP and extrapolate to 10²⁶ FLOPs. The final number: about $9 billion trillion in present-day dollars to train a Claude Opus equivalent in 1946. That’s 70 million times more than current world GDP. Even if Shannon had wanted to buy this much computing power, it was beyond infeasible for the economy of his time.

Before comparing that figure to today’s cost to build Claude Opus, let’s take a pitstop in 2012, right at the cusp of the deep learning revolution. In the halls of the University of Toronto, Geoffrey Hinton and two soon-to-be-famous graduate students are training what will eventually be known as AlexNet – a pioneering image recognition model which bet on scaling compute as a way to yield better performance. How much would it have cost Hinton to train Claude Opus instead of AlexNet?

We’ll again use the largest supercomputer in the world at the time for reference. The TITAN supercomputer had just been built at a cost of $97 million. It was capable of 2.7 x 10¹⁶ FLOPs per second and drew 8200 kW of power. Using a similar calculation with estimates for power draw and cost, it would cost $12 billion in present-day dollars to train Claude with many TITANs. Given the improvements in hardware efficiency from ENIAC to TITAN, training Claude in 2012 would have been hypothetically feasible for a nation-state or megacap company, even if the cost was still beyond what would have been accessible to Hinton and collaborators.

Today? With the same FLOP extrapolation on contemporary compute, it would cost around $105 million to train Claude Opus in 2026. Note the cost reductions: from Shannon to Hinton, we see a 719 trillion-fold reduction in cost. But from 2012 to the present, we see a mere 114-fold reduction. By these stylized examples, efficiency increases from Shannon to Hinton were more than 6 trillion times more than from Hinton to the present.

Across the same time periods, the growth in compute investment rather than hardware efficiency is far less dramatic. Across notable machine learning models, the investment growth from 1946 to 2012 is only around 600x more than the growth from 2012 to the present, on average.3 Hence, from Shannon to Hinton, we see almost 2,000x more growth in hardware efficiency than investment. But from Hinton to today, we see more than 6,000,000x more growth in investment than hardware efficiency.

This divergence is consequential: the last decade has seen tremendous investment with minimal efficiency gains, but rides on the back of 60+ years of efficiency accumulation – billions upon trillions of times more efficient hardware. While current capabilities are in part the result of rapid growth in compute buildout by frontier labs, they stand on the shoulders of a half-century of Moore’s law.

The next ten years of progress matter a lot

We’ve seen that historically, the slow march of hardware efficiency has played a much larger role than the recent boom of compute investment. What does that mean for the future of AI progress? We see roughly three possible universes.

In the first universe, true transformative AI is only a few orders of magnitude away. Frontier labs can simply reach up and grab it with enough short-term spending, provided they can find investors who can front the cash without requiring immediate returns. In this world, hardware progress over the last 70 years has taken us just short of the end zone, and we can reap the full rewards with one last monumental push.

However, training compute cannot increase indefinitely without world-historical economic returns. A naive extrapolation of current spending would have frontier model training compute spending exceed world GDP by 2040. It’s not plausible that there is enough patient liquid investment capital to reach this point. So if we don’t see transformative AI in the next several orders of magnitude of scaling, that brings us to the second universe.

In the second universe, transformative AI doesn’t arrive soon, and those capital constraints mean that hardware investment must slow to match the pace of economic returns. However, capabilities improvements do bring in enough revenue that investment still proceeds at a relatively rapid pace, and progress is still driven mostly by hardware investment. The pace of progress is faster than what we’d see from hardware efficiency alone.

In the final universe, the next breakthrough is again many orders of magnitude away, and the utility of training more expensive models diminishes fast enough that scaling cannot pay for itself. Companies cannot find a way to economically continue exponential scaling, and we are instead forced to wait decades for hardware progress to slowly accumulate. Just like Claude Opus was simply out of reach for Claude Shannon, it may be that no short-term investment can substitute for the gradual progress of decades.

The next few years are critical for determining which of these paths we are on. If the CEOs of frontier labs are correct, we could see transformative AI in the next few years as the labs make their final push. Short of such a breakthrough, we might still see that demand grows so rapidly that frontier labs need not slow their buildout spend, suggesting that the flywheel is fast enough to compensate for slow hardware progress. But if capabilities aren’t enough to drive revenue growth, spending will slow tremendously, and we, like Shannon, will be forced to watch the clock and ride the gradual wave of efficiency improvements for decades.

Subscribe now

This post was written by Zach Brown and Alex Fogelson. Neil Thompson provided editorial support. Lucy Yan provided logistical support.

Appendix: calculation details

ENIAC: To calculate the cost of training Claude using ENIAC, we use its true lifecycle of 9 years, an initial build-out cost of $487,000, a max power draw of 150 kW, an estimated power cost of $0.02 per kWh, and FLOP rate of 500 FLOPS. The cost per FLOP is assuming it operates at maximum FLOPS and power capacity for 9 years, and the final cost is converted to 2026 dollars with a 16.93 inflation multiplier. We don’t estimate the utilization for ENIAC. Our final number is about 9 billion trillion dollars in 2026.

TITAN: To calculate the cost of training Claude using TITAN, we use both the initial cost of construction and the cost of a one time upgrade, totalling $97 million, amortized over its seven year lifespan. For power draw, we estimate a cost per kWh of $0.06 and a max power draw of 8200 kW. The peak FLOPS for TITAN was 27e15, though we assume 25% utilization, which was deemed historically reasonable by Claude Opus and GPT 5.5. Again adjusting for 2026 dollars, we get $12 billion dollars:

Claude, Today: Finally, for our modern day Claude estimations, we use an approximate cost per hour of renting an H100 of $1.50, a utilization of 40%, and a peak FLOP rate of 990 TFLOPS to get $105 million.

We here at FutureTech remain cautiously ambivalent about the magnitude of these claims. But AI’s transformational impact on science is directionally correct, and we need not prophesize to claim this. For scientists, AI enabled methodologies are neither a recent innovation nor a punctuated transition. Quite the contrary: the use of AI models across science has been growing near-exponentially for well over a decade, across nearly all domains of science.

In our new paper The Rapid Growth of AI Foundation Model Usage in Science, we document this historical trend with unprecedented depth, tracking who has adopted which foundation models and how scientists deploy them in their methodologies. We construct the largest dataset of its kind – classifying 750 thousand academic citations of nearly 2 thousand foundation models across 250 thousand papers. In this post, we’ll not only demonstrate the staggering growth of AI in science over the last decade, but we’ll look under the hood at which models are used, how that impacts scientific outcomes, and why it matters. Expect to learn:

1. Which fields are currently using AI foundation models most often, growing that use the fastest, and using the largest models?

2. How quickly are scientists keeping up with the AI frontier? How do research outcomes change as scientists move closer to the frontier?

3. What bottlenecks in AI may be limiting scientific breakthroughs of the sort predicted by frontier labs?

To begin, we’ll set the stage with a fundamental question: what is an AI foundation model, and how do we track their deployment in science?

What do we mean by “AI foundation models”?

The term “AI” means different things to different people. Before we can track how scientists are using AI, we need to agree on some definitions. Is a chess engine like IBM’s DeepBlue – hardcoded with domain knowledge, deterministically navigating through search space – considered AI? What about a simple linear regression model, fit using gradient-based methods? Are we just being obnoxiously pedantic? Perhaps, like Justice Potter Stewart, we’ll just know it when we see it!

Sadly, we won’t. The literature has no standard definition of what “AI” means in the context of scientific methodologies. For instance, this paper uses a list of keywords to indicate the presence of AI, but they seemingly don’t publish that list anywhere. This report from Stanford uses an elaborate ontology to classify titles and abstracts, but eventually uses word embeddings – an uninterpretable way of defining AI statistically – to approximate AI topic engagement. This study even includes support vector machines, which are hardly more deserving of the “AI” label than drawing a line through a scatter plot.

We take a more concrete approach. Rather than asking about anything which could be AI, we focus more narrowly on scientists using specific existing AI models such as AlexNet, BERT, or GPT-4. These so-called foundation models are deep learning models which required high upfront costs at the time of training (e.g. compute, data, expertise), but are general enough to be (cheaply) reused for a variety of scientific tasks. Our dataset is constructed as follows:

1. We first curate a list of foundation models, a comprehensive set of their metadata, and the papers which introduced them.1 This allows us to track more than the mere presence of AI, including model size, modality, openness, age, institution of origin, and much more.

2. We use academic citation graphs to track which publications cite papers that introduced a foundation model. For each citation, we find the exact context of the citation, while also extracting comprehensive metadata on the citing paper. This provides a more accurate signal for AI engagement than paper abstracts or keywords, while still being compute efficient.

3. Finally, each time a foundation model is cited, we classify the citation context into one of three categories: background (just mentioning the model), usage (taking the model as-is, off the shelf), and extending (adapting/fine-tuning the model). This classification is essential when leveraging citation network patterns: counting all citations would overestimate true adoption rates by 141%.

Together, this dataset lets us answer questions about specific patterns of model adoption in unprecedented detail. Which fields are using the largest models? The oldest models? Which fields require the most fine-tuning? Which countries and institutions dominate among adopters?

In this post, we will focus on differences between scientific domains, and save questions about who is adopting foundation models for later posts. Enough with the methodological minutia, let’s get to the results!

All Fields See Near-Exponential Growth

From a bird’s eye view, we see that the adoption2 of foundation models across all of science has been growing near-exponentially for almost a decade. By 2024, more than 1% of all academic publications adopted a foundation model, with around one-third actively customizing (extending). The share of adopters in the literature has grown more than 33x over the last decade.

This trend isn’t restricted to a few strands of the literature either. We see staggering growth across all domains, from Computer Science and Engineering to Linguistics and Psychology. Among these fields, Linguistics (34%) and Computer Science (18%) stand out with the higher adoption rates, perhaps unsurprisingly.3 But nearly every field in our analysis reached 1% adoption by 2024 – domains like History and Sociology included – highlighting the diverse potential of foundation models across disciplines.

Many fields with lower adoption rates still show astonishing growth rates. Biology and Chemistry exhibit the fastest growth for model usage (off-the-shelf adoption) with 3-year compound growth rates of 123% and 168%, respectively. Both of these fields also exhibit strong growth in model customization, but Psychology is racing ahead of the pack with a 107% annual growth rate. (See Table 1 for a more complete breakdown.)

This extraordinary growth doesn’t tell the full story. Scientists are using AI more now than ever before, but which models are they using? When we analyze the properties of models being adopted by researchers, we don’t find the same story of rapid growth.

Bigger Does in Fact Mean Better

In AI, bigger models are more costly to use. But bigger models are also better models. That’s not just a rule of thumb: the size of a model (and training dataset) predicts its performance with extraordinary accuracy according to neural scaling laws.4 That predictability is great news if you are a frontier lab detonating your longest ever training run. But for compute constrained scientists, scaling laws have a different take home message.

Neural scaling laws are power laws of compute with very small exponents. These power laws change very, very slowly, such that smaller and smaller performance improvements require larger and larger models. A few GPUs might be helpful for running smaller models, but if you need frontier performance, those GPUs may quickly become as useful as a TI-84.

Scientists are adopting foundation models more than ever, but given the importance of model size for performance, are those models also getting bigger? We observe that the size of adopted models is growing, but it’s growing slowly – far slower than the size of models being built at the frontier. In 2013, the median model built was 7.7x larger than the median model adopted. By 2024, that number jumped to 26x. The models adopted by scientists aren’t keeping pace with exponentially increasing sizes at the frontier.

Some fields are doing better than others, though. Psychology, Linguistics, and Law lead the pack in 2024 as the only three fields adopting (on average) models with more than 3 billion parameters.5 Computer Science sits in the middle, while Materials Science, Environmental Science, and Engineering bring up the rear under 100 million parameters.

Perhaps fields like Psychology or Law needed LLMs before AI could be of substantial use for most applications, whereas Materials Science or Engineering can still find use in small models at the AlexNet scale. This explanation is supported by the contribution of language models to recent growth in adopted model sizes. From 2022 to 2024, average size of models adopted by scientists increased by 3.9x. But if we restrict adoptions to only language and multimodal models, sizes increase by 5x and 11x, respectively.

This lag also exists when measuring the age of the models. The average age of models adopted (at the time of use) is at least 3.3 years old across all fields. Linguistics and Law use the newest models (3.3 and 3.8 years old, respectively), which is consistent with the hypothesis presented above. Perhaps surprisingly, Mathematics is far less cutting-edge (5.7 years), despite modern LLMs abilities in competitive mathematics and novel research. Though this could be from lagging diffusion, mathematics may also be especially prone to leveraging frontier models without citation. Bringing up the rear is Physics (5.9 years) and Materials Science (6.1 years), straying far from the cutting edge, but consistent with the hypothesis that certain fields can still squeeze value out of older, smaller models.

Of course, these differences only matter if newer, bigger models translate into better scientific outcomes in practice, not just in theory. And indeed, this is exactly what we find (though only correlationally). On average, since 2020, increasing the size of an adopted foundation model 10x corresponds to 1.3x more citations. That same 10x also corresponds to higher-quality journals, with an average increase of 2.3 points in journal impact factor.6

We suspect using larger models yields better performance, better research, and therefore better publication outcomes. Of course, it’s possible the causality could go in the opposite direction: teams that publish the best papers may accrue more computing resources and be free to use larger models. Or there could be some third factor: for example, maybe working in industry leads to both better outcomes and more resources. But in each of these, the correlation we find is still suggestive. Why do researchers with ample compute choose to use bigger models, which still take more time and resources to run? Why would industry allocate precious compute resources to research if it weren’t necessary for better results?

Scientists Are Clearly Compute Constrained

Our findings suggest that the best science is done with the biggest models. But bigger models also require more computing power. Sadly, academic science is notably constrained by insufficient computing resources. In fact, the “compute divide” between public and private science is growing rapidly.

Observing the top 500 AI supercomputers across the world, we see that the share of public ownership has fallen rapidly, from 61% in 2019 to 12% in 2025. Public ownership went from a healthy majority to a miniscule fraction over barely half a decade. Infrastructure investment looks even bleaker. The most notable public sector investment is the NSF’s recent National Artificial Intelligence Research Resource (NAIRR) pilot program, which by mid-2024 had delivered a total compute capacity equal to 5,000 H100 GPUs. That’s a great start, but 5,000 distributed across many universities is missing a few zeros from top industry counts like Microsoft (150,000), Meta (150,000), Google (50,000), or Amazon (50,000).

At MIT FutureTech, we’ve focused on this issue for years. We directly observed this compute divide by comparing the sizes and shares of foundation models built between industry and academia. We’ve documented that AI-intensive research is more dependent on capital – like computing resources – than traditional research (which depends more on human labor). Now this most recent paper shows that AI is becoming a cornerstone of modern science, but outcomes may be compute constrained. We believe that scientific research – particularly non-proprietary discoveries – are an essential source of economic growth and progress. To fully reap those benefits, scientists will need enough resources to discover the next generation of breakthroughs.

Subscribe now

This post was written by Alex Fogelson. Ana Trišović, Zachary Brown, Sebastian Sartor, and Neil Thompson provided editorial feedback. Lucy Yan provided logistic support. The authors of the paper are Ana Trišović, Alex Fogelson, Janakan Sivaloganathan, and Neil Thompson. For correspondence on the paper, please contact Ana Trišović or Neil Thompson.

We recently had the chance to tackle such questions during our workshop on the Economics of AI and Automation. The workshop brought together economists studying what new models reveal about the pace of automation and economic change. Discussion throughout the day examined how firms test and adopt new AI capabilities, how financial markets react to model releases, and how these pieces of information shape our expectations for AI-induced growth and risk. In this post, we’ll distill a few points from those discussions:

1. Demand for new models rarely comes from cross-creator substitution (for example: when Anthropic releases Claude Opus 4.5, demand mostly isn’t drawn from existing Gemini users);

2. Firm automation decisions are driven by the outsized fixed costs of AI;

3. Financial markets are not predicting imminent transformative AI.

Demand for new models rarely comes from cross-creator substitution

Andrey Fradkin and his coauthors show that markets for language models are dynamic and competitive, with leading models dominating for a few months before being displaced. However, substitution patterns vary across providers and fall into two main types: some models cannibalize incumbents from the same creator, while others expand overall demand. Using data on model downloads and prices on OpenRouter, a platform for AI models, he shows that the release of Claude Sonnet 3.7 (for example) shifted demand sharply away from Claude Sonnet 3.5, an instance of within-creator substitution. But there was little to no contemporaneous decline in demand for competing models from other creators. Other releases, such as Gemini 2.5 Pro and xAI’s Grok Code Fast 1, show rapid adoption with minimal displacement of other models. This might suggest that these models increased aggregate demand, rather than reallocating existing demand.

It surprised us to see older models often retain users when newer models come out, even though OpenRouter makes it very seamless to swap models. Perhaps this is an indication that experimentation is expensive, making switching costs larger than they might appear at first glance; or that users have built their applications around a particular model that make switching difficult. Or it could simply be evidence of differentiation in the market for language models.

Firm automation decisions are driven by the outsized fixed costs of AI

Three papers throughout the workshop tackled the role of fixed costs in the development and impact of AI models. Rebekah Dix and coauthors show that cloud computing technology can turn what was traditionally a high fixed cost into a variable cost. Still, high fixed costs to procure computing infrastructure are still essential for understanding the economics of AI adoption: an MIT FutureTech paper (presented by Danial Lashkari and written with fellow FutureTech members Wensu Li, Neil Thompson and Christina Qiu) argues that fixed costs in many areas slow the pace of automation and shape its resultant impact on the economy. In a third presentation, Flavio Calvino and Luca Fontanelli confirm that many firms still find it worthwhile to pay the fixed cost of AI development.

In most technology-intensive industries, firms must pay some sort of fixed cost to get started. For example, a bank aiming to replace its customer service representatives with an AI chatbot must first fine-tune a model on its private data, and procure some quantity of computing infrastructure to serve the chatbot – both fixed costs that the firm pays up-front. Because firms can only imperfectly predict the future, these up-front investments can generate inefficiencies. If more customers than anticipated use the chatbot, the firm might have difficulty serving the chatbot to all of them with its stock of computing infrastructure. Dix et al. show how on-demand cloud computing providers like AWS solve this problem for the bank. Using cloud computing, it can quickly scale up and down its compute stock to adjust to varying demand. In other words, cloud computing turns a fixed cost into a variable cost.

But, as Lashkari and coauthors discuss, AI adoption still involves a number steps that incur significant fixed costs. Returning to our example, if only a few customers use a bank’s chatbot, it may not be worth it for the bank to pay the fixed costs of the fine-tuning process for so little return. The widespread presence of fixed costs such as thislike this one affects which tasks are worthwhile to automate. If a firm only needs to accomplish a few instances of a task, it might not be worthwhile to pay the up-front cost of investment in AI. This mechanism means that some tasks might not be automated even if frontier AI is more efficient than humans on a per-task basis.

Still, Calvino and Fontanelli show that many firms do adopt and develop AI despite its fixed cost. They demonstrate clear differences between firms that choose to purchase AI models from external developers (AI buyers), and those that choose to develop them in-house (AI developers). Using detailed survey data from French firms, they show that AI developers tend to be larger firms, but also newer. They also emphasize the role of complementary assets, showing that developers hold stock of IT assets. Speaking directly to the question of fixed costs, Flavio argues that the development of AI systems may involve higher fixed costs than purchasing but that firms which develop their AI tools in-house see a productivity boost from AI. The up-front cost might be high, but it pays off.

Markets are not predicting imminent transformative AI

Basil Halperin and co-authors argue that long-horizon real interest rates might provide a signal of AI progress. In standard economic models, any meaningful increase in the probability of transformative AI should raise long-term real interest rates by reducing the value of future consumption, regardless of whether AI yields explosive productivity growth or existential catastrophe. Why? The real interest rate is the cost of borrowing resources from the future to spend today. If transformative AI means we are either going to be enjoying the bounty of a plentiful world centered around aligned AI or we are heading to the abyss – either way – why wait to consume tomorrow? Carpe diem! Consume more today! In this framework, elevated long run rates would indicate markets pricing in a sharp break from historical growth paths, regardless of whether the anticipated outcome is overwhelmingly positive or negative. Of course, these assumptions assume that markets are informed and pricing risk efficiently.

Maryam Farboodi, an MIT FutureTech affiliate, along with her fellow MIT co-author Isaiah Andrews, provided convincing data that, as of September 2025, markets are not acting as though they expect transformative AI to be just around the corner. Studying bond markets around major AI model releases, she finds that long-dated yields consistently fall following announcements. This suggests that, conditional on new information from a release, markets revise beliefs away from transformation rather than toward it. Taken together, these results imply that markets are paying attention to AI progress, but the results also suggest that markets interpret recent releases as lowering growth expectations, (or, following Halperin et al., interpret them as evidence of a reduced likelihood of AI-induced catastrophe). Model announcements appear to shift expectations toward delayed or more modest impact — a conclusion that stands in contrast to the more dramatic narratives that often dominate in the immediate aftermath of a release.

Why might markets not be expecting higher growth from AI? One new idea came from Joshua Gans, who presented a framework examining how AI affects the direction of research. He models AI as an interpolation technology that fills gaps between existing areas of research. The key insight is that AI’s impact on growth depends on a critical threshold in its capabilities. When AI has modest interpolation range, connecting areas of research that already are “close together,” it encourages incremental research that increases the density of knowledge in an area of research, which – Gans argues – raises growth in a world where ideas are getting harder to find. But when AI becomes sufficiently capable, it might promote exploratory research that connects far-apart research areas, reducing knowledge density, paradoxically slowing growth in the long-run despite productivity gains in the short-run.

Concluding

The workshop itself was a great success, and we are grateful to all the presenters and participants who contributed to such a rich set of discussions. A fascinating set of seven talks by Rebekah Dix, Flavio Calvino, Maryam Farboodi, Andrey Fradkin, Basil Halperin, Joshua Gans and Danial Lashkari highlighted how much progress has been made in understanding AI adoption from both macro and micro perspectives, and also how much remains an open question.

Building on this momentum, MIT FutureTech has more research in the pipeline on related topics. For instance, an upcoming paper, “The Birth, Life, and Death of Models”, describes what happens when a model is released, how it is used throughout its “life”, and how it eventually fades away from economic relevance. That is not all for this academic year: we will be hosting another round of our Economics of AI and Automation workshop in the spring, continuing the conversation as both the technology and the evidence evolve.

Subscribe now

Joseph Emmens, Kazimier Smith, Zach Brown, and Won-Jae Chang wrote this post. Aaron Kaye, Alex Fogelson, Sebastian Sartor, Zach Brown, and Neil Thompson provided editorial support. Lucy Yan provided logistical support.

The sources of progress in AI capabilities are typically broken down into three distinct components:

• Compute scaling counts the raw number of mathematical operations, or FLOPs, performed during model training. More FLOPs mean better performance. You can perform more total operations when you (1) have more total hardware or (2) run that hardware for longer.

• Hardware efficiency quantifies the costs of each operation in monetary, energy, or time units. Better hardware efficiency implies more FLOPs can be performed for the same number of dollars, joules, or days.

• Algorithmic efficiency encompasses the cleverness of the training procedure, quantifying the performance “bang” for each FLOP “buck.” Better algorithmic efficiency means you can achieve the same model performance while using fewer operations.

Typically, each of these components is treated as orthogonal to the others, such that improvements can be evaluated independently between sources of progress. For example, hardware efficiency (in FLOPs per dollar) improved 45x between 2013 and 2024. Because compute scale and hardware efficiency are considered independent, you can expect your hardware efficiency gains to be roughly that same 45x regardless of whether you have 10 GPUs or 10,000.1

Traditionally, algorithmic efficiency is also thought to be independent of compute scale. Whether you have 10¹⁵ FLOPs or 10²⁵ FLOPs, the literature assumes that when researchers discover better training techniques, applying these algorithmic improvements should always yield the same multiplicative efficiency improvements – regardless of compute scale (e.g. 10x less compute at 10¹⁵ FLOPs and 10²⁵ FLOPs alike). But is this a safe assumption?

Our new paper provides strong experimental evidence against this view. We find that algorithmic efficiency is compute scale-dependent: some algorithms provide increasing – not constant – efficiency returns to scale. We estimate that from 2012 to 2025, as much as 91% of efficiency gains depend on (historically exponential) compute scaling, meaning that if compute scaling had not occurred, algorithmic innovations alone would yield less than 10% of the efficiency improvements we measure today.

Our paper further examines the downstream implications of these findings while laying the theoretical foundations for analyzing algorithmic progress in light of this scale-dependence. In this post, we’ll cover:

1. How to quantify algorithmic progress in AI;

2. The stunning role of scale-dependence in this progress, as revealed by our experiments;

3. What questions remain to inform future research.

How do we quantify algorithmic progress in AI pre-training?

If you are familiar with the notion of an algorithm’s time-complexity, you’ve probably heard of Big-O notation, which gives us a concise language for capturing the asymptotic features of classical algorithms. Stating an algorithm has complexity O(n²) means you’ll need something like “the square of the input size” steps before terminating. Specifically, this notation is ideal when we care about runtime across varying input sizes, not the precise number of steps.

For AI algorithms, we need a different framework for comparison. Unlike traditional algorithms, in machine learning, a training algorithm doesn’t produce a correct answer from a fixed-size input. Instead, it utilizes some quantity of FLOPs to output a model with a given performance level. The relationship between FLOPs and performance are known as neural scaling laws, and they will help us define a framework for measuring algorithmic efficiency in machine learning. To help illustrate, we’ll call on our ever-reliable friends Alice and Bob.

Imagine Alice and Bob have two different training algorithms. Bob is given some fixed compute budget of 10²⁰ FLOPs to use with his algorithm, and creates a model with some performance level. To analyze the efficiency improvements from Alice’s algorithm (or lack thereof), we can ask “how much more/less compute would Alice’s scaling law require to reach the same level of performance that Bob’s model achieved with 10²⁰ FLOPs?” Perhaps Alice has a better algorithm which only needs 10¹⁹ FLOPs to reach performance parity with Bob’s model. In this case, we would say Alice’s algorithm achieves a 10x compute-equivalent gain (CEG) relative to Bob’s 10²⁰ FLOPs model.

In the existing literature, including work from our lab, this CEG multiplier of 10x is assumed to hold no matter Bob’s compute budgets. We would assume, for example, that if Bob trained a 10²⁵ FLOPs model, it would match Alice’s 10²⁴ FLOPs model in performance. But in our paper, we show empirically that–just like the time complexity of a classical algorithm–the efficiency gain from an algorithmic innovation may depend on the compute scale. For example, it’s possible that when Bob uses 10¹⁵ FLOPs, Alice’s algorithm uses 10x less compute (10¹⁴ FLOPs) to reach performance parity; but when Bob uses 10²⁰ FLOPs, Alice can get away with using 100x less (only 10¹⁸ instead of 10¹⁹ FLOPs). Efficiency gains are compute scale-dependent.

Using an expanded framework which allows for this scale-dependence, our paper further breaks down the algorithmic innovations of the past decade into those which do and don’t depend on compute scale – and quantifies how much each matters. We show that most progress has indeed come from steadily increasing compute budgets, which makes Alice’s algorithm appear to be more and more efficient by comparison to Bob’s. This is in contrast to the dominant view in the literature, where progress comes from steady innovations in Alice’s algorithm.

In the next section, we’ll outline the empirical results and methods that lead to this finding.

Scale-dependence (and a lot of compute) is all you need!

To rigorously test the efficiency gains from different algorithmic innovations, we use validation loss as our performance metric, which measures how well the model approximates the training data’s distribution. We ran two kinds of experiments: ablation experiments and scaling experiments.

• Ablation experiments surgically remove targeted innovations from the training algorithms to measure their performance contributions. This isolates the role of specific algorithms.

• Scaling experiments measure performance for different algorithms across different compute scales, deriving scaling laws in the process. This isolates the role of compute scale for a particular training algorithm.

Most of the algorithms we tested (e.g. optimizers, normalization, encodings) were scale-invariant changes. That is, they improved efficiency by a constant factor regardless of the size of the model trained. How do we know this? We tested a cluster of algorithms over a range of compute scales.2 When removing all of these algorithms at once, the change detected in our scaling law was the same percentage of compute across all compute scales.3 This provides strong evidence that these changes individually have scale-invariant efficiency improvements.4

However, two of the innovations we explored were strongly scale-dependent:

1. the transition from LSTM to Transformer architectures, and

2. the transition to Chinchilla-optimal scaling.

These innovations represented the overwhelming share of net efficiency improvements. Running scaling experiments between LSTMs and Transformer separately, we show that this architecture switch has scale-dependent efficiency improvements.5 As the compute budget for the LSTM increases, the potential compute savings of switching to a Transformer get larger and larger.

We also analytically explored changes in data-parameter balancing: when training large models, an inevitable tradeoff between the model size and the dataset size determines the total compute cost. Over time, our understanding of the optimal balance has improved. Kaplan et al. recommended tilting more heavily towards parameters, while later work from Hoffman et al. (known as “Chinchilla scaling”) advocated for a more balanced approach. We show that switching from Kaplan’s once state-of-the-art recommendation to the Chinchilla scaling yields increasing efficiency returns with compute scale. In other words, the cost of balancing incorrectly gets worse and worse with more compute.

Although our experiments are done at small scales, we can extrapolate our findings to estimate the efficiency improvements across larger compute budgets. Research done by Epoch AI suggests that frontier compute levels have followed a pretty smooth exponential growth trend over time. Using these estimates, we create a detailed image of efficiency improvements over historical compute scales:

1. We recover 6,930x of the 22,000x improvement in algorithmic progress estimated by our previous paper with Epoch AI from 2012 through 2023. Though naively it may seem like we only capture one-third of total algorithmic progress, leaving 3x out of 22,000x unaccounted for means we account for more than 88% of total multiplicative efficiency gains.

2. We find that at 2025 frontier compute levels, 91% of measured algorithmic progress came from scale-dependent innovations. These gains would not have been reaped without increasing compute budgets, meaning models trained with less compute have seen less than 10% of this efficiency improvement.

Below you can see the precise breakdown of progress decomposed into specific innovations. The innovations are depicted starting when they were discovered, and the compute scale is determined by the trend of notable models in that year. The scale-invariant changes, despite their quantity, represent a small fraction of the efficiency gains by 2025. The large blue region represents the switch to Transformers, while the lighter red region represents Chinchilla rebalancing.

If models trained with less compute aren’t getting the same benefits, that has implications for the environment, science, and society more broadly. Smaller models use orders of magnitude less energy, and therefore fewer emissions.6 Smaller foundation models are the dominant option for scientific methodologies across disciplines. Smaller models are essential for private, local AI on personal devices. The figure below highlights the stark differences in efficiency gains across model compute scales. Models trained with 10¹⁶ FLOPs see less than 30x efficiency gains, while models trained with 10²⁴ FLOPs reap more than 20,000x improvement! These are implications we hope to explore in future research.

How could we improve measures of algorithmic progress?

The framework we’ve proposed is an extension of the CEG framework from the existing literature. However, the notion of a compute equivalent gain admits two important limitations, both of which point to important future research on how we conceptualize algorithmic progress. In particular, CEG is (1) highly reference frame-dependent and (2) relies on choosing the right model performance metric.

Reference-Frame Dependence

Reference-frame dependence is a counterintuitive property that we elucidate in depth in the paper, and which only arises when efficiency improvements are scale-dependent. Importantly, this phenomenon does not occur for the kinds of efficiencies people usually reason about in everyday life, which are scale-independent.

To see the contrast, consider a familiar case where scale dependence is absent. Suppose you buy a new iPhone 16 and are told its battery efficiency is 10x better than the first iPhone. Then you hear that the iPhone 17 is 100x more efficient than the first iPhone. In that setting, it would be perfectly reasonable—and indeed unavoidable—to conclude that the iPhone 17 is about 10x more battery-efficient than the iPhone 16. Battery efficiency improvements compose transitively and can be compared without worrying about reference frames.

Algorithmic efficiency, as measured by compute-equivalent gains (CEGs), behaves differently when improvements are scale-dependent. In particular, it is possible to have a sequence of models trained with steadily increasing compute (e.g. progressively larger, otherwise unchanged Transformer models) where each model is no more compute-efficient than the previous one. Yet relative to a different reference algorithm, such as an LSTM, the same sequence can appear to exhibit continual efficiency gains. In this case, nothing about the training algorithm itself has improved; instead, increasing compute moves the model along a scaling curve where the gap to the reference algorithm’s scaling curve grows.

This creates the appearance of algorithmic progress without new innovations. Without explicitly running scaling and ablation experiments of the sort we do in the paper, simply observing improved performance at larger scales can lead the CEG framework to attribute efficiency gains to “algorithmic progress” that are in fact inherited from scale-dependent properties of earlier innovations.

Performance Metrics

When measuring model performance, we use cross-entropy loss as our metric, which is standard in the scaling law literature. It tells you how good your prediction is relative to the true data distribution. This subtly makes our performance metric “how well does my model fit the training data?” Algorithmic progress, however, encompasses more than the fitness of the model to the data.

Modern chatbots, for example, don’t want to approximate the exact pre-training distribution. Through reinforcement learning, large language models subtly augment their output distribution to mimic instruction following, obey safety guardrails, or emphasize chain-of-thought reasoning. Measuring performance in those domains is less straightforward, and is not well captured by the CEG framework without a clearly defined benchmark or fitness metric.

Moreover, many recent algorithmic innovations haven’t focused on pre-training at all. Instead, a large chunk of computing costs come from serving models, making the cost of inference also economically important. Lots of innovations focus on improving efficiency after the model is trained, with techniques like distillation and quantization targeting runtime efficiency and performance tradeoffs. (See here for some recent work from our lab on the pace of progress in these areas.)

Wrapping up

At MIT FutureTech, algorithmic progress has been a pillar of our research for many years, well beyond algorithms for machine learning. We’ve built the world’s largest database of algorithms, created a framework and a tool for assessing the advantage of quantum algorithms, applied this tool to chemistry and deep learning, and estimated the rate of progress in large language models. Across each of these domains, analyzing progress requires leveraging the right framework to clearly capture the salient innovations of different kinds of algorithms.

Broadly speaking, we think analyzing algorithmic efficiency is a multifaceted problem which still requires new ideas and frameworks to properly capture. Models perform across many axes and their costs are multidimensional. To fully capture the efficiency of a model, we need better ways of measuring capabilities, training inputs, and pre-training/inference tradeoffs.

You can read the paper here. Stay subscribed to Mixture of Experts to get the latest updates about these research questions.

Subscribe now

This post was written by Alex Fogelson. Zachary Brown, Sebastian Sartor, Lucy Yan, Hans Gundlach, and Neil Thompson provided editorial feedback. The authors of the paper are Hans Gundlach, Alex Fogelson, Jayson Lynch, Ana Trišović, Jonathan Rosenfeld, Anmol Sandhu, and Neil Thompson. For correspondence on the paper, please contact Hans Gundlach or Neil Thompson.

That’s what Audrey, a high school student from New York (with, apparently, quite short timelines), wrote in to the New York Times in February. She went on:

“Employers in the future will be more likely to choose robots that work full time, don’t take lunch breaks, are almost always right, and don’t need to be paid, over humans who need to be paid, require breaks, complain, and sometimes make errors in their work. I am fearful that by the time I go to college and am ready to start a real job, there will be nowhere for me to work [...] I need to do my research and go into a field that is not vulnerable to automation.”

Meanwhile Robert, a student at West Chicago High School, is more sanguine:

“In the past, major advancements like the Industrial Revolution or the rise of computers disrupted employment, yet new opportunities followed. While some jobs will vanish due to AI, this shift will likely create new roles and industries. The real challenge isn’t avoiding change — it’s adapting to it like we always have.”

Similar conversations are ripping through schools, board rooms, research labs, and government offices around the world. Will AI automation happen quickly or slowly? What forces will govern the transition? Will AI augment or displace human labor? And what will the consequences be: for the economy, for scientific progress, and for Audrey and Robert?

Audrey, perceptively, cited the importance of research. At the MIT FutureTech conference in October, entitled How Fast Will AI Automation Happen?, we brought together some of the leading figures studying AI development, the economics of AI, and AI applications to discuss these topics. This post highlights a selection of ideas presenters raised. It’s not a full summary of each talk, just connections we found interesting. We’ve loosely organized the post around three themes:

1. Technical and economic constraints to AI-fueled explosive growth,

2. AI adoption and the labor market, and

3. AI automation in science.

We’ll take these in turn. YouTube links to the talks are included throughout.

Technical and economic constraints to AI-fueled explosive growth

The rate and nature of economic growth from AI depends, of course, on the pace of AI capability improvements. Many speakers referenced the astonishing progress language models have made on benchmarks, such as METR’s time horizon benchmark, a particularly impressive example. But, in many cases, speakers also tempered expectations of rapid improvement.

For one thing, the potential for capabilities improvements often varies by domain and type of AI model. In some cases, there may be room for AI to strongly outperform humans, but in other domains, we may already be close to optimal performance. MIT FutureTech’s Neil Thompson gave the example of AlphaChip, a Google AI tool that helps optimize its semiconductor designs. The latest generation of AlphaChip achieved a 6.2% improvement over their baseline performance, an impressive but relatively modest improvement. By contrast, Google was able to achieve a 40% reduction in datacenter cooling costs using machine learning many years early, a far more substantial reduction.

What should we make of progress in language models and general purpose AI? Gary Marcus, characteristically skeptical, argued that large language models still possess significant limitations. While capable at tasks within their training distribution, Marcus suggests they cannot effectively generalize beyond the data they have seen. He also notes that they continue to suffer from decades-old problems, like hallucinations, poor planning skill, and inconsistent world models. While benchmark performance has been impressive, Tamara Broderick questioned whether commonly-used metrics of AI capabilities are robustly capturing what we care about, pointing out that some widely-discussed LLM leaderboard rankings are extremely sensitive to a few data points. Anna Li highlighted that memory miniaturization already hit a scaling wall many years ago, and continued scaling of LLMs for long-context tasks risks becoming bottlenecked by high bandwidth memory constraints. She argued that we will need algorithmic improvements, such as quantization and RL, which reduce the memory footprint without relying on hardware advances.

One much discussed potential issue with widespread AI use is the energy and climate cost of training and running models. Elsa Olivetti walked us through the basic climate economics of AI use, demonstrating that a dramatic increase in AI use would need to either increase its share of energy consumption (already 4.4% in the US as of 2023), bidding up the price for all uses, or expand power production (as many frontier companies plan to1). In addition to the climate burden from energy use, producing the GPUs themselves leaves a large carbon footprint – Olivetti told us this was about 320kg of CO2e per GPU. Whether this factor constrains AI datacenter build-outs or AI induced growth, though, may depend on policy responses.

Beyond technical bottlenecks, economic bottlenecks could constrain the impacts of AI. One of Ben Jones’s slides stuck with us: over the last fifty years, computing power has improved by a factor of something like a hundred thousand trillions. Yet despite the deep integration of computing technology across the economy, the total size of the economy has only grown by a factor of about three. In economic parlance, there are tasks or production processes which are complementary to automated sectors, but which themselves are not feasible to automate. Even if dramatic efficiency improvements in automated sectors mean hugely expanding the production of some goods, the sectors that do not see efficiency improvements may constrain total output. In her talk, Beena Ammanath described some of these bottlenecks in business contexts, further emphasizing the practical difficulties of fully integrating AI.

AI adoption and the labor market

Even if AI capabilities progress is rapid, it can take time for AI to become integrated into production processes2, and measuring adoption during this transition can be challenging. How can we best understand the current role of AI in the workforce? And as AI diffuses, how should we expect automation to impact labor and wages?

Kristina McElharan showed empirical data on AI adoption pre-ChatGPT, finding that only about six percent of firms used machine learning anywhere in their production process. Yet this figure is much higher for large firms: in her sample, 62% of large firms were using machine learning somewhere in their business. Together, these figures highlight two important points: AI adoption can be slow and heterogenous, and measuring the rate of adoption depends on methodological choices. Going forward, we’re taking this as a reminder to scrutinize AI adoption statistics: what was the sample, and what precisely was measured?

In another talk, David Deming presented results from a more recent sample weighted to be representative of the American public according to the US Current Population Survey. Deming and his coauthors find that AI adoption is quite impressive, with about 40% of workers responding that they used generative AI tools like ChatGPT in the past week, either at work or at home, and about 27% used AI at work specifically. Deming and his coauthors also polled workers on the time savings they attribute to generative AI, finding that each hour spent using generative AI increases a worker’s productivity by about 33%. Given current usage rates, this implies that generative AI tools could be boosting US productivity by as much as two percent. In another study, he and his coauthors examine ChatGPT usage data from OpenAI, finding that usage rates have been increasing over time not only from new AI adopters, but from more intense usage of ChatGPT by existing adopters.

Several speakers mentioned the “shadow” use of AI – AI used by workers informally, and perhaps without the knowledge of managers. This could explain why the numbers Deming and his coauthors find when surveying individuals is more than most of the numbers McElharan and coauthors find when surveying firms. Indeed Deming finds some evidence for this phenomenon: among users on the consumer plan (not even the enterprise plan) about a third of messages seem to be related to work tasks. Importantly, this shadow deployment may not take full advantage of the AI productivity gains from coordinated, firm-wide deployment.

In another talk, Frank Nagle suggested that the presence of cheap, open-source models can make AI more attractive to adopt and help AI diffuse more quickly. He showed that the gap between open and closed model capabilities is narrowing quickly: while it used to take open models about between ten and forty weeks or so to catch up to closed-model performance, recently the gap has shrunk to ~between ten and fifteen weeks. Nagle also presented evidence that AI inference costs have gone down dramatically. Leveraging this data, he hypothesized that open-source models have effectively capped the price of closed-source models, which have seen cheaper inference costs despite increasing capabilities. (These declining inference costs not only increase the pace of adoption, but also reduce the cost of test-time compute scaling and of automating AI research, either of which could accelerate AI capabilities.)

Altogether, the speakers presented a mixed view of AI adoption rates: AI adoption is increasing, and its pace is impressive compared to past technologies, but adoption is not uniform. Nor is it always clear that AI is being implemented in a meaningful and deliberate way by firms. But one might naturally ask: if AI adoption is pervasive enough to automate a large number of worker tasks, how will that affect workers?

Traditional models of automation assume that the first-order effects of automation are bad for workers: when a job’s task is automated, there will be less employment in that job and lower wages for workers who perform it. Countervailing effects may occur as the economy adjusts, such as an expansion of output leading to increased demand for workers (an instance of “Jevons Paradox”). And workers who lose their jobs to automation (or some of their job tasks) can sometimes move into new, higher paying roles – maybe even ones that are created by the automating technology in the first place. Still, these adjustments are second order consequences of automation.

However, David Autor presented a more nuanced picture in a presentation on his recent paper “Expertise” (written with MIT FutureTech’s Neil Thompson with research assistance from FutureTech’s Lucy Hampton). To understand the argument, suppose jobs are made up of a bundle of different tasks, each requiring different levels of expertise. When automation occurs, it affects jobs differently depending on the relative expertise of a task within an occupation. If more expert tasks are automated, the job becomes less specialized overall, increasing the supply of eligible workers and therefore lowering wages. However, if less expert tasks are automated, the job becomes more specialized, and wages increase. Automation is therefore context dependent, and automating the same task can have different impacts on different occupations.

AI automation in science

Full automation of scientific research and development would be transformative, likely leading to substantial increases in the rate of economic growth. This is because science is a durable source of economic growth ordinarily; when automated, the returns to science could be continually reinvested into more AIs doing science. Indeed, Ben Jones has argued that, while automating ordinary production of goods can lead to a “Type I” growth explosion, where growth rates increase without bound but remain finite at any timepoint, only automating the production of ideas can lead to a “Type II” growth explosion – a singularity – where models show the economy achieving infinite output in finite time.

So, given the potential transformative impact of automating science: what does AI automation in science look like today? Several of our speakers discussed the integration of AI into existing scientific research practice. For instance, Karianne Bergen spoke about AI use in the earth sciences. Here, machine learning models can “emulate” computationally expensive models of complex earth systems, like weather systems, predicting the output of the expensive model. This reduces time and cost for downstream researchers. Shantanu Singh discussed efforts to automate and scale research in medicine and lab sciences. For example, in drug testing, AI models can be trained to identify phenotypic features of healthy and diseased cells, automating the detection of phenotypic markers from treated cells and allowing researchers to test many candidate molecules or doses simultaneously. He also emphasized that AI coding assistants and agents can prove useful in a wide range of research tasks, especially in data analysis. Patrick Ruch discussed the use of AI tools in scientific documentation and in research, including materials science, chemistry, and algorithms. Yilun Du discussed some nascent methods for training robot foundation models using synthetic data from video generation models. Though these results remain exploratory and preliminary, the technique he presented would provide a way of generating robotic training data for any scientific discipline, so long as video generation models could coherently produce a high-fidelity replication of the task at hand. Since physical tasks remain a bottleneck in the laboratory sciences, it is especially important to track improvements in embodied AI.

But even if we automate substantial portions of existing scientific production pipelines, remaining unautomated areas will prove to be bottlenecks eventually. Yutaro Yamada described Sakana AI’s attempts to avoid this issue by building an automated scientific production pipeline from scratch. Their agent stitches together a number of language models which read academic papers, generate hypotheses, perform computational experiments to test those hypotheses, interpret the results, and generate academic papers. They have also experimented with a “meta-agent” which learns how best to design such an agent. One paper authored by this system was successfully published at the International Conference on Learning Representations (ICLR).

However, open questions remain about how well their approach can generalize to less computational areas of science, about the risks of models learning to “game” the peer review process, and about the type of science that approaches like this produce. In particular, Neil Thompson mentioned that it may not be enough to automate the “extensive margin” of science – that is, to simply scale up the amount of experimentation we do using existing methods or to apply existing research in a straightforward way. Rather, the most impressive science is on the “intensive margin”: it does things better, not just more. Scientific progress requires an expansion of the distribution of scientific ideas, correcting the mistakes in existing research and addressing previously unknown gaps.

Moving forward

These conversations highlighted just how many unknowns remain. We are still early in this transition, and many of the mechanisms shaping AI’s impact on science, workers, and society remain mysterious. There are more open questions than settled conclusions. We think research is more important now than ever.

A heartfelt thank you to all presenters for sharing their expertise, and to all attendees for contributing to a thoughtful discussion.

Thanks as well to the conference organizers: Emanuele Del Sozzo, Matthias Mertens, Rob Seamans, and Neil Thompson for organizing talks; Tess Fagan, Sam Brown, Lucy Yan and Shiobhain Jenkins for operations; the audiovisual team and catering staff; and the FutureTech team who assisted with operations during the conference.

This post was written by Zachary Brown. Alex Fogelson, Sebastian Sartor, Lucy Yan, and Neil Thompson provided editorial feedback. Lucy Yan helped plan this post and organize the talk links.

Who We Are

MIT FutureTech is an interdisciplinary research lab jointly based at MIT CSAIL and the MIT Sloan School of Management. Founded in 2019 by Dr. Neil Thompson, the lab brings together more than a hundred researchers working at the intersection of technology, economics, and policy.

Our mission is to understand and quantify the causes and consequences of progress in AI and computing through rigorous, high-impact research. This requires a mixture of experts: we integrate expertise across hardware, software, algorithms, scaling laws, AI, robotics, macroeconomics, industrial organization, labor markets, and much more. Our resulting insights are leveraged by decision-makers in business, government, and society, as well as technical researchers and practitioners.

What We Study

At MIT FutureTech, we study trends in computing power (e.g. hardware innovation, Moore’s Law), how that computing power is used (AI development, algorithms, software, science), and how those uses will impact the broader society (e.g. AI automation, AI risks, and AI in R&D).

For example, in our 2020 paper “The Computational Limits of Deep Learning,” we showed that deep learning’s rapid progress is highly dependent on increases in raw computing power. We argued that without major improvements in efficiency, this trajectory will become economically and environmentally unsustainable. In our 2022 paper “The Importance of (Exponentially More) Compute” we provided evidence of similar dynamics across a number of historically important computing bellwethers (like chess) and economically important problems (like weather prediction). Across domains, progress in capabilities has required exponential increases in computing power that will become increasingly expensive to maintain.

If it is increasingly infeasible to maintain exponential growth in raw compute capacity, can algorithms and efficiency improvements compensate? To what degree? We’ve studied this question in traditional computer science problems and in deep learning contexts. In deep learning, we found that improving algorithms for language models can double our compute efficiency in less than a year, substantially outpacing hardware gains. Beyond these paradigms, we’ve even explored new computing paradigms such as quantum computing, analyzing where and when quantum systems may achieve a competitive edge over classical architectures based on their algorithmic and economic differences, and deploying these insights in our Quantum Economic Advantage Calculator.

The remarkable progress in both computation and algorithms has led to remarkable advancements in AI capabilities. We want to better understand the economic impacts of these advancements. In “Beyond AI Exposure: Which Tasks Are Cost-Effective to Automate with Computer Vision?”, we showed that many economically valuable tasks are not profitable to automate even when it is technically feasible to do so. When automation does occur, we’ve argued that an important consideration is the expertise required for humans to perform the remaining unautomated tasks: if a job sees its most expert tasks automated, the skill required to perform it decreases, and wages go down as more people enter the profession; when inexpert tasks are automated, the job requires a greater degree of skill, and wages rise as employment falls.

When automation occurs in the production of science itself, the effects can be profound. In “Economic Impacts of AI-augmented R&D”, we explored how AI accelerates scientific idea generation, demonstrating that capital-intensive AI systems can dramatically boost research productivity and economic growth. Our ongoing work measures the extent to which AI use in science is already occurring, with indications that usage is growing exponentially.

Progress in AI has also brought about new risks. To support a more coordinated and evidence-based discussion, we created the MIT FutureTech AI Risk Repository, a living database that organizes 777 documented risks drawn from 43 taxonomies. This open resource provides a shared frame of reference for researchers, policymakers, and practitioners to better understand and address the spectrum of AI risks.

Looking Ahead

The work presented here represents a fraction of the work we do. Our ongoing research at FutureTech explores the role of open-source models, market dynamics of building and deploying large foundation models, deep dives into AI risks, and much more.

Through this Substack, we’ll share insights into our ongoing projects, our published research, and commentary on emerging developments in computing, AI, and the economics of technological change.

We believe these are some of the most important topics of our time. We hope you’ll join us in the discussion.

Subscribe now

This piece was written jointly by Alex Fogelson, Sebastian Sartor, and Zachary Brown. Lucy Yan was essential for planning this Substack, and provided excellent operational and communications support. Thanks to Hans Gundlach for the blog name, Lucy Hampton for her expertise, and Neil Thompson for reviewing.